Add Qwen1.5-MoE Q8_0 CPU and single-token TornadoVM GPU inference - #144

Merged

mikepapadim merged 10 commits into

beehive-lab:mainfrom

MRPRESIDENT66:moe-single-token

Aug 6, 2026

Commits on Aug 5, 2026

Add Qwen1.5-MoE (qwen2moe) CPU inference support
MRPRESIDENT66
committed
97ef9d2
Fix Qwen2MoEState to override createStateFields with correct config type
MRPRESIDENT66
committed
46cea94
Fix Qwen2-MoE routing probabilities
MRPRESIDENT66
committed
973cdf3
Add ChatML turn separator for Qwen messages
MRPRESIDENT66
committed
8ddca7d
Optimize Q8_0 dot products with quantized activations
MRPRESIDENT66
committed
25dd61b
Clean up Qwen2-MoE CPU implementation
MRPRESIDENT66
committed
71d49d3
Add Qwen2-MoE single-token GPU baseline
MRPRESIDENT66
committed
c2e9583
Update baseline branch name
MRPRESIDENT66
committed
2e843c6
chore(qwen2-moe): remove local validation artifacts
MRPRESIDENT66
committed
b227bc2

Commits on Aug 6, 2026

Fix Qwen2-MoE RMSNorm reduction race
MRPRESIDENT66
committed
7a4f197